In [ ]:
#Lets see how this images look;
plt.imshow(train_images[1050], interpolation='nearest')
plt.title('Infected = ' + str(train_labels[1050]))
plt.show()
The context: Why is this problem important to solve?
This problem is important to solve since malaria is a disease that affects millions each year, many with resulting death.
Having a system that can detect malaria early and fastly can reduce mortality and therefore save many lives.
The objectives: What is the intended goal?
The intended goal is to develop an deep learning model to detect parasitized cells from microscopic images
The key questions: What are the key questions that need to be answered?
The key questions that need to be answer are:
How does the data look and what kind of transformations need to be done in order to work with it?
Which model architecture is best suited for this data?
And what measurements are we going to use in order to get such model?
The problem formulation: What is it that we are trying to solve using data science?
Simply put, we are trying to generate a deep learning model that classifies images microscopic images from cells into parasitized or uninfected, which implies that the organisms has malaria.
There are a total of 24,958 train and 2,600 test images (colored) that we have taken from microscopic images. These images are of the following categories:
Parasitized: The parasitized cells contain the Plasmodium parasite which causes malaria
Uninfected: The uninfected cells are free of the Plasmodium parasites
from google.colab import drive
drive.mount('/content/drive')
Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D, BatchNormalization, Activation, LeakyReLU
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import backend
from keras.callbacks import EarlyStopping, ReduceLROnPlateau, ModelCheckpoint, Callback
import zipfile
import cv2
import os
import random
from PIL import Image
from sklearn.metrics import classification_report
from sklearn.metrics import confusion_matrix
print("Tensorflow version " + tf.__version__)
#try:
# tpu = tf.distribute.cluster_resolver.TPUClusterResolver() # TPU detection
# print('Running on TPU ', tpu.cluster_spec().as_dict()['worker'])
#except ValueError:
# raise BaseException('ERROR: Not connected to a TPU runtime; please see the previous cell in this notebook for instructions!')
#tf.config.experimental_connect_to_cluster(tpu)
#tf.tpu.experimental.initialize_tpu_system(tpu)
#tpu_strategy = tf.distribute.TPUStrategy(tpu)
Tensorflow version 2.12.0
path = '/content/drive/My Drive/Colab Notebooks/cell_images.zip'
#The data is provided as a zip file so we need to extract the files from the zip file
with zipfile.ZipFile(path, 'r') as zip_ref:
zip_ref.extractall()
for f in zip_ref.namelist():
zinfo = zip_ref.getinfo(f)
if(zinfo.is_dir()):
print(f)
cell_images/test/ cell_images/test/parasitized/ cell_images/test/uninfected/ cell_images/train/ cell_images/train/parasitized/ cell_images/train/uninfected/
# I define a function to extract data from each directory, as an image of a given size
def get_data(dir, status, size):
images = []
labels = []
for x in os.listdir(dir):
path = dir + x
img = cv2.imread(path, cv2.IMREAD_UNCHANGED)
img = cv2.resize(img, (size,size))
images.append(np.array(img))
labels.append(status)
return images, labels
The extracted folder has different folders for train and test data will contain the different sizes of images for parasitized and uninfected cells within the respective folder name.
The size of all images must be the same and should be converted to 4D arrays so that they can be used as an input for the convolutional neural network. Also, we need to create the labels for both types of images to be able to train and test the model.
Let's do the same for the training data first and then we will use the same code for the test data as well.
#Image size parameter
size = 80
train_1, label_1 = get_data('/content/cell_images/train/parasitized/', 1, size)
train_2, label_2 = get_data('/content/cell_images/train/uninfected/', 0, size)
train_images = np.array(train_1 + train_2)
train_labels = np.array(label_1+label_2)
test_1, tlabel_1 = get_data('/content/cell_images/test/parasitized/', 1, size)
test_2, tlabel_2 = get_data('/content/cell_images/test/uninfected/', 0, size)
test_images = np.array(test_1+ test_2)
test_labels = np.array(tlabel_1+tlabel_2)
del train_1
del train_2
del test_1
del test_2
del label_1
del label_2
del tlabel_1
del tlabel_2
print('Shape of train data:', train_images.shape)
print('Shape of train label data:', train_labels.shape)
print('Shape of test data:', test_images.shape)
print('Shape of test label data:', test_labels.shape)
Shape of train data: (24958, 80, 80, 3) Shape of train label data: (24958,) Shape of test data: (2600, 80, 80, 3) Shape of test label data: (2600,)
#Lets see how this images look;
plt.imshow(train_images[1050], interpolation='nearest')
plt.title('Infected = ' + str(train_labels[1050]))
plt.show()
print('Shape of train images: ', train_images[0].shape)
print('Shape of test images: ',test_images[0].shape)
Shape of train images: (80, 80, 3) Shape of test images: (80, 80, 3)
print(train_labels[1])
print(test_labels[1])
1 1
The data is ready to be manipulated.
print('Max pixel value for train data:', train_images.max())
print('Min pixel value for train data:', train_images.min())
print('Max pixel value for test data:', test_images.max())
print('Min pixel value for test data:', train_images.min())
Max pixel value for train data: 255 Min pixel value for train data: 0 Max pixel value for test data: 255 Min pixel value for test data: 0
Everithing seems fine.
print('Number of parasitized in train data:',len(train_labels[train_labels==1]))
print('Number of uninfected in train data:',len(train_labels[train_labels==0]))
print('Number of parasitized in test data:',len(test_labels[test_labels==1]))
print('Number of uninfected in test data:', len(test_labels[test_labels==0]))
Number of parasitized in train data: 12582 Number of uninfected in train data: 12376 Number of parasitized in test data: 1300 Number of uninfected in test data: 1300
x_train_normalized = train_images/255
x_test_normalized = test_images/255
fig = plt.figure(figsize=(12,12))
ax = fig.add_subplot(2,1,1)
ax.set_title('Uninfected and Parasitized in Train Data')
aux1 = pd.DataFrame(train_labels)
sns.countplot(aux1, x = 0,ax = ax)
ax = fig.add_subplot(2,1,2)
ax.set_title('Uninfected and Parasitized in Test Data')
aux2 = pd.DataFrame(test_labels)
sns.countplot(aux2, x = 0,ax = ax)
plt.show()
print('Values for train data:\n', aux1.value_counts())
print('Values for test data:\n',aux2.value_counts())
Values for train data: 1 12582 0 12376 dtype: int64 Values for test data: 0 1300 1 1300 dtype: int64
It looks like the data is balanced.
Let's visualize the images from the train data
#I create a function to visualize the images since I'm going to use it later.
#It takes x= image data, y = label date, number of rows and cols, and the indexes to be shown
def plot_data(x,y, rows, cols, k):
fig = plt.figure(figsize = (12, 12))
for i in range(cols):
for j in range(rows):
ax = fig.add_subplot(rows, cols, i * rows + j + 1)
ax.imshow(x[k[i * rows + j ]])
ax.set_title('Infected = ' + str(y[k[ i * rows + j ]]) + '\n Obs #:' + str(k[ i * rows + j ]))
plt.show()
#Here I choose random indexes for train and test data. I compute them because later
#I would like to seee what happens to the same images under some transformations
np.random.seed(37)
rows = 6
cols = 6
k1 = []
k2 = []
for x in range(36):
k1.append(np.random.randint(train_images.shape[0]))
k2.append(np.random.randint(test_images.shape[0]))
print('Train Data')
plot_data(train_images, train_labels, rows, cols, k1)
Train Data
print('Test Data')
plot_data(test_images, test_labels, rows, cols, k2)
Test Data
It looks like infected cells have some areas tinted that differentiate them.
aux1 = np.rint(np.mean(train_images[train_labels == 1], axis = 0)).astype(int)
aux0 = np.rint(np.mean(train_images[train_labels == 0], axis = 0)).astype(int)
fig = plt.figure(figsize = (12, 12))
ax = fig.add_subplot(2, 1, 1)
ax.imshow(aux1)
ax.set_title('Average image of Parasitized for train data ')
ax = fig.add_subplot(2, 1, 2)
ax.imshow(aux0)
ax.set_title('Average image of Uninfected for train data ')
plt.show()
Both images look very similar, with the average for infected a little darker.
train_hgv = np.empty((train_images.shape[0], size,size,3), dtype=int)
for x in range(train_images.shape[0]):
train_hgv[x] = cv2.cvtColor(train_images[x], cv2.COLOR_RGB2HSV)
test_hgv = np.empty((test_images.shape[0], size,size,3), dtype=int)
for x in range(test_images.shape[0]):
test_hgv[x] = cv2.cvtColor(test_images[x], cv2.COLOR_RGB2HSV)
plot_data(train_hgv, train_labels, rows, cols,k1)
plot_data(test_hgv, test_labels, rows, cols, k2)
del train_hgv
del test_hgv
#I am liberating some memory
train_blurred = np.empty((train_images.shape[0], size,size,3), dtype=int)
for x in range(train_images.shape[0]):
train_blurred[x] = cv2.GaussianBlur(train_images[x], (5,5),0)
test_blurred = np.empty((test_images.shape[0], size,size,3), dtype=int)
for x in range(test_images.shape[0]):
test_blurred[x] = cv2.GaussianBlur(test_images[x], (5,5),0)
plot_data(train_blurred, train_labels, rows, cols, k1)
plot_data(test_blurred, test_labels, rows, cols, k2)
del train_blurred
del test_blurred
#liberating some memory
Gaussian blurring removes noise from images. It can help if we are sure that is a problem. However, we don't know wich kernel / signma to use, and using a gaussian blur with fixed might be dominated by using a gaussian filter in the CNN.
For this particular problem it seems that it won't be very helpful.
Note: The Base Model has been fully built and evaluated with all outputs shown to give an idea about the process of the creation and evaluation of the performance of a CNN architecture. A similar process can be followed in iterating to build better-performing CNN architectures.
#imported at the beginning
#parameters for the models
epoch = 15
val = 0.2
batch = 300
y_train_encoded = tf.keras.utils.to_categorical(train_labels,2)
y_test_encoded = tf.keras.utils.to_categorical(test_labels,2)
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
def model_1():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 5, padding = "same", input_shape = (size, size, 3), activation= 'relu'))
model.add(MaxPooling2D(pool_size = 2))
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_1 = model_1()
print(model_1.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 1216
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
flatten (Flatten) (None, 25600) 0
dense (Dense) (None, 32) 819232
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 820,514
Trainable params: 820,514
Non-trainable params: 0
_________________________________________________________________
None
Fit and train our Model
es = EarlyStopping(monitor = 'val_loss', mode = 'min', verbose = 1, patience = 5)
mc = ModelCheckpoint('best_model.h5', monitor = 'val_accuracy', mode = 'max', verbose = 1, save_best_only = True)
history_1 = model_1.fit(
x_train_normalized, y_train_encoded,
epochs = epoch,
validation_split = val,
shuffle = True,
batch_size=batch,
use_multiprocessing = True,
callbacks = [es, mc]
)
Epoch 1/15 67/67 [==============================] - ETA: 0s - loss: 0.6844 - accuracy: 0.5984 Epoch 1: val_accuracy improved from -inf to 0.00000, saving model to best_model.h5 67/67 [==============================] - 6s 27ms/step - loss: 0.6844 - accuracy: 0.5984 - val_loss: 0.9430 - val_accuracy: 0.0000e+00 Epoch 2/15 66/67 [============================>.] - ETA: 0s - loss: 0.6022 - accuracy: 0.6332 Epoch 2: val_accuracy improved from 0.00000 to 0.15986, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.6021 - accuracy: 0.6335 - val_loss: 0.9880 - val_accuracy: 0.1599 Epoch 3/15 66/67 [============================>.] - ETA: 0s - loss: 0.5851 - accuracy: 0.6871 Epoch 3: val_accuracy improved from 0.15986 to 0.20413, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.5850 - accuracy: 0.6873 - val_loss: 0.9396 - val_accuracy: 0.2041 Epoch 4/15 66/67 [============================>.] - ETA: 0s - loss: 0.5752 - accuracy: 0.7025 Epoch 4: val_accuracy improved from 0.20413 to 0.25260, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.5754 - accuracy: 0.7017 - val_loss: 0.9881 - val_accuracy: 0.2526 Epoch 5/15 66/67 [============================>.] - ETA: 0s - loss: 0.5720 - accuracy: 0.7131 Epoch 5: val_accuracy improved from 0.25260 to 0.31651, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.5717 - accuracy: 0.7126 - val_loss: 0.9739 - val_accuracy: 0.3165 Epoch 6/15 66/67 [============================>.] - ETA: 0s - loss: 0.5637 - accuracy: 0.7242 Epoch 6: val_accuracy improved from 0.31651 to 0.49539, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.5638 - accuracy: 0.7241 - val_loss: 0.8883 - val_accuracy: 0.4954 Epoch 7/15 66/67 [============================>.] - ETA: 0s - loss: 0.5582 - accuracy: 0.7298 Epoch 7: val_accuracy improved from 0.49539 to 0.51783, saving model to best_model.h5 67/67 [==============================] - 1s 14ms/step - loss: 0.5582 - accuracy: 0.7293 - val_loss: 0.8558 - val_accuracy: 0.5178 Epoch 8/15 66/67 [============================>.] - ETA: 0s - loss: 0.5551 - accuracy: 0.7314 Epoch 8: val_accuracy did not improve from 0.51783 67/67 [==============================] - 1s 14ms/step - loss: 0.5551 - accuracy: 0.7318 - val_loss: 1.0053 - val_accuracy: 0.3095 Epoch 9/15 66/67 [============================>.] - ETA: 0s - loss: 0.5465 - accuracy: 0.7388 Epoch 9: val_accuracy did not improve from 0.51783 67/67 [==============================] - 1s 14ms/step - loss: 0.5466 - accuracy: 0.7384 - val_loss: 0.8755 - val_accuracy: 0.4639 Epoch 10/15 66/67 [============================>.] - ETA: 0s - loss: 0.5410 - accuracy: 0.7452 Epoch 10: val_accuracy did not improve from 0.51783 67/67 [==============================] - 1s 13ms/step - loss: 0.5412 - accuracy: 0.7449 - val_loss: 0.9259 - val_accuracy: 0.3846 Epoch 11/15 65/67 [============================>.] - ETA: 0s - loss: 0.5343 - accuracy: 0.7496 Epoch 11: val_accuracy did not improve from 0.51783 67/67 [==============================] - 1s 13ms/step - loss: 0.5345 - accuracy: 0.7502 - val_loss: 1.0358 - val_accuracy: 0.2582 Epoch 12/15 66/67 [============================>.] - ETA: 0s - loss: 0.5277 - accuracy: 0.7570 Epoch 12: val_accuracy did not improve from 0.51783 67/67 [==============================] - 1s 13ms/step - loss: 0.5274 - accuracy: 0.7570 - val_loss: 0.9216 - val_accuracy: 0.4916 Epoch 12: early stopping
#I create functions to look at the performance of the models
def evaltest(x_test, model):
y_pred_test = model.predict(x_test)
y_pred_test_classes = np.argmax(y_pred_test, axis = 1)
y_pred_test_max_probas = np.max(y_pred_test, axis = 1)
return y_pred_test, y_pred_test_classes, y_pred_test_max_probas
def class_conf(y_test, y_pred):
print(classification_report(y_test, y_pred))
# Plotting the heatmap using confusion matrix
cm = confusion_matrix(y_test, y_pred)
plt.figure(figsize = (8, 5))
sns.heatmap(cm, annot = True, fmt = '.0f')
plt.ylabel('Actual')
plt.xlabel('Predicted')
plt.show()
def trainvalcurve(history_1):
plt.plot(history_1.history['accuracy'])
plt.plot(history_1.history['val_accuracy'])
plt.title('Model Accuracy')
plt.ylabel('Accuracy')
plt.xlabel('Epoch')
plt.legend(['Train', 'Validation'], loc = 'upper left')
plt.show()
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(x_test_normalized, model_1)
82/82 [==============================] - 0s 2ms/step
Plotting the confusion matrix
class_conf(test_labels, y_pred_test_classes)
precision recall f1-score support
0 0.78 0.48 0.60 1300
1 0.63 0.87 0.73 1300
accuracy 0.67 2600
macro avg 0.70 0.67 0.66 2600
weighted avg 0.70 0.67 0.66 2600
Plotting the train and validation curves
trainvalcurve(history_1)
So now let's try to build another model with few more add on layers and try to check if we can try to improve the model. Therefore try to build a model by adding few layers if required and altering the activation functions.
backend.clear_session()
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
def model_2():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 3, padding = "same", input_shape = (size, size, 3), activation= 'relu'))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 32, kernel_size = 5, padding = 'same', activation= 'sigmoid'))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 64, kernel_size = 3, padding = 'same', activation= 'relu'))
model.add(MaxPooling2D(pool_size = 2))
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_2 = model_2()
print(model_2.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 448
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
conv2d_1 (Conv2D) (None, 40, 40, 32) 12832
max_pooling2d_1 (MaxPooling (None, 20, 20, 32) 0
2D)
conv2d_2 (Conv2D) (None, 20, 20, 64) 18496
max_pooling2d_2 (MaxPooling (None, 10, 10, 64) 0
2D)
flatten (Flatten) (None, 6400) 0
dense (Dense) (None, 32) 204832
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 236,674
Trainable params: 236,674
Non-trainable params: 0
_________________________________________________________________
None
Using Callbacks
es = EarlyStopping(monitor = 'val_loss', mode = 'min', verbose = 1, patience = 5)
mc = ModelCheckpoint('best_model.h5', monitor = 'val_accuracy', mode = 'max', verbose = 1, save_best_only = True)
Fit and Train the model
history_2 = model_2.fit(
x_train_normalized, y_train_encoded,
epochs = epoch,
validation_split = val,
shuffle = True,
batch_size=batch,
use_multiprocessing = True,
callbacks = [es, mc]
)
Epoch 1/15 67/67 [==============================] - ETA: 0s - loss: 0.6628 - accuracy: 0.6268 Epoch 1: val_accuracy improved from -inf to 0.00000, saving model to best_model.h5 67/67 [==============================] - 4s 32ms/step - loss: 0.6628 - accuracy: 0.6268 - val_loss: 0.9881 - val_accuracy: 0.0000e+00 Epoch 2/15 65/67 [============================>.] - ETA: 0s - loss: 0.6371 - accuracy: 0.6298 Epoch 2: val_accuracy did not improve from 0.00000 67/67 [==============================] - 1s 16ms/step - loss: 0.6367 - accuracy: 0.6302 - val_loss: 0.9856 - val_accuracy: 0.0000e+00 Epoch 3/15 66/67 [============================>.] - ETA: 0s - loss: 0.6136 - accuracy: 0.6480 Epoch 3: val_accuracy improved from 0.00000 to 0.49439, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.6136 - accuracy: 0.6478 - val_loss: 0.8194 - val_accuracy: 0.4944 Epoch 4/15 65/67 [============================>.] - ETA: 0s - loss: 0.5979 - accuracy: 0.6789 Epoch 4: val_accuracy did not improve from 0.49439 67/67 [==============================] - 1s 16ms/step - loss: 0.5974 - accuracy: 0.6790 - val_loss: 0.9619 - val_accuracy: 0.2732 Epoch 5/15 66/67 [============================>.] - ETA: 0s - loss: 0.5888 - accuracy: 0.6909 Epoch 5: val_accuracy improved from 0.49439 to 0.51502, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.5885 - accuracy: 0.6910 - val_loss: 0.8337 - val_accuracy: 0.5150 Epoch 6/15 65/67 [============================>.] - ETA: 0s - loss: 0.5749 - accuracy: 0.7061 Epoch 6: val_accuracy improved from 0.51502 to 0.55749, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.5750 - accuracy: 0.7058 - val_loss: 0.8196 - val_accuracy: 0.5575 Epoch 7/15 67/67 [==============================] - ETA: 0s - loss: 0.5664 - accuracy: 0.7136 Epoch 7: val_accuracy improved from 0.55749 to 0.64583, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.5664 - accuracy: 0.7136 - val_loss: 0.7324 - val_accuracy: 0.6458 Epoch 8/15 65/67 [============================>.] - ETA: 0s - loss: 0.5599 - accuracy: 0.7196 Epoch 8: val_accuracy did not improve from 0.64583 67/67 [==============================] - 1s 15ms/step - loss: 0.5599 - accuracy: 0.7197 - val_loss: 0.9146 - val_accuracy: 0.4878 Epoch 9/15 66/67 [============================>.] - ETA: 0s - loss: 0.5487 - accuracy: 0.7338 Epoch 9: val_accuracy did not improve from 0.64583 67/67 [==============================] - 1s 15ms/step - loss: 0.5487 - accuracy: 0.7337 - val_loss: 0.8396 - val_accuracy: 0.5343 Epoch 10/15 67/67 [==============================] - ETA: 0s - loss: 0.5391 - accuracy: 0.7395 Epoch 10: val_accuracy did not improve from 0.64583 67/67 [==============================] - 1s 16ms/step - loss: 0.5391 - accuracy: 0.7395 - val_loss: 0.7607 - val_accuracy: 0.5911 Epoch 11/15 66/67 [============================>.] - ETA: 0s - loss: 0.5261 - accuracy: 0.7488 Epoch 11: val_accuracy did not improve from 0.64583 67/67 [==============================] - 1s 15ms/step - loss: 0.5264 - accuracy: 0.7485 - val_loss: 0.9678 - val_accuracy: 0.3638 Epoch 12/15 63/67 [===========================>..] - ETA: 0s - loss: 0.5111 - accuracy: 0.7566 Epoch 12: val_accuracy improved from 0.64583 to 0.66446, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.5103 - accuracy: 0.7568 - val_loss: 0.6941 - val_accuracy: 0.6645 Epoch 13/15 65/67 [============================>.] - ETA: 0s - loss: 0.4800 - accuracy: 0.7823 Epoch 13: val_accuracy did not improve from 0.66446 67/67 [==============================] - 1s 16ms/step - loss: 0.4802 - accuracy: 0.7813 - val_loss: 0.7224 - val_accuracy: 0.6146 Epoch 14/15 65/67 [============================>.] - ETA: 0s - loss: 0.4488 - accuracy: 0.7973 Epoch 14: val_accuracy improved from 0.66446 to 0.75741, saving model to best_model.h5 67/67 [==============================] - 1s 16ms/step - loss: 0.4484 - accuracy: 0.7976 - val_loss: 0.5529 - val_accuracy: 0.7574 Epoch 15/15 67/67 [==============================] - ETA: 0s - loss: 0.4015 - accuracy: 0.8325 Epoch 15: val_accuracy did not improve from 0.75741 67/67 [==============================] - 1s 16ms/step - loss: 0.4015 - accuracy: 0.8325 - val_loss: 0.6356 - val_accuracy: 0.6831
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(x_test_normalized, model_2)
82/82 [==============================] - 0s 3ms/step
Plotting the confusion matrix
class_conf(test_labels, y_pred_test_classes)
precision recall f1-score support
0 0.87 0.62 0.72 1300
1 0.70 0.91 0.79 1300
accuracy 0.76 2600
macro avg 0.79 0.76 0.76 2600
weighted avg 0.79 0.76 0.76 2600
Plotting the train and the validation curves
trainvalcurve(history_2)
Now let's build a model with LeakyRelu as the activation function
Let us try to build a model using BatchNormalization and using LeakyRelu as our activation function.
backend.clear_session()
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
def model_3():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 3, padding = "same", input_shape = (size, size, 3)))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 32, kernel_size = 3, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(BatchNormalization())
model.add(Conv2D(filters = 64, kernel_size = 5, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Dropout(0.3))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_3 = model_3()
print(model_3.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 448
leaky_re_lu (LeakyReLU) (None, 80, 80, 16) 0
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
conv2d_1 (Conv2D) (None, 40, 40, 32) 4640
leaky_re_lu_1 (LeakyReLU) (None, 40, 40, 32) 0
max_pooling2d_1 (MaxPooling (None, 20, 20, 32) 0
2D)
batch_normalization (BatchN (None, 20, 20, 32) 128
ormalization)
conv2d_2 (Conv2D) (None, 20, 20, 64) 51264
leaky_re_lu_2 (LeakyReLU) (None, 20, 20, 64) 0
max_pooling2d_2 (MaxPooling (None, 10, 10, 64) 0
2D)
dropout (Dropout) (None, 10, 10, 64) 0
batch_normalization_1 (Batc (None, 10, 10, 64) 256
hNormalization)
flatten (Flatten) (None, 6400) 0
dense (Dense) (None, 32) 204832
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 261,634
Trainable params: 261,442
Non-trainable params: 192
_________________________________________________________________
None
Using callbacks
es = EarlyStopping(monitor = 'val_loss', mode = 'min', verbose = 1, patience = 5)
mc = ModelCheckpoint('best_model.h5', monitor = 'val_loss', mode = 'min', verbose = 1, save_best_only = True)
Fit and train the model
history_3 = model_3.fit(
x_train_normalized, y_train_encoded,
epochs = epoch,
validation_split = val,
shuffle = True,
batch_size=batch,
use_multiprocessing = True,
callbacks = [es, mc]
)
Epoch 1/15 67/67 [==============================] - ETA: 0s - loss: 0.6644 - accuracy: 0.6246 Epoch 1: val_loss improved from inf to 0.69490, saving model to best_model.h5 67/67 [==============================] - 5s 33ms/step - loss: 0.6644 - accuracy: 0.6246 - val_loss: 0.6949 - val_accuracy: 6.0096e-04 Epoch 2/15 65/67 [============================>.] - ETA: 0s - loss: 0.5890 - accuracy: 0.6378 Epoch 2: val_loss improved from 0.69490 to 0.69244, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.5884 - accuracy: 0.6387 - val_loss: 0.6924 - val_accuracy: 0.9776 Epoch 3/15 65/67 [============================>.] - ETA: 0s - loss: 0.5537 - accuracy: 0.7034 Epoch 3: val_loss improved from 0.69244 to 0.68739, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.5536 - accuracy: 0.7031 - val_loss: 0.6874 - val_accuracy: 0.9894 Epoch 4/15 67/67 [==============================] - ETA: 0s - loss: 0.4615 - accuracy: 0.7847 Epoch 4: val_loss improved from 0.68739 to 0.32143, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.4615 - accuracy: 0.7847 - val_loss: 0.3214 - val_accuracy: 1.0000 Epoch 5/15 64/67 [===========================>..] - ETA: 0s - loss: 0.3294 - accuracy: 0.8669 Epoch 5: val_loss improved from 0.32143 to 0.08843, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.3264 - accuracy: 0.8682 - val_loss: 0.0884 - val_accuracy: 1.0000 Epoch 6/15 65/67 [============================>.] - ETA: 0s - loss: 0.2368 - accuracy: 0.9117 Epoch 6: val_loss improved from 0.08843 to 0.03806, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.2361 - accuracy: 0.9122 - val_loss: 0.0381 - val_accuracy: 1.0000 Epoch 7/15 65/67 [============================>.] - ETA: 0s - loss: 0.1799 - accuracy: 0.9338 Epoch 7: val_loss improved from 0.03806 to 0.01925, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.1799 - accuracy: 0.9337 - val_loss: 0.0193 - val_accuracy: 1.0000 Epoch 8/15 65/67 [============================>.] - ETA: 0s - loss: 0.1501 - accuracy: 0.9450 Epoch 8: val_loss improved from 0.01925 to 0.01389, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.1498 - accuracy: 0.9452 - val_loss: 0.0139 - val_accuracy: 1.0000 Epoch 9/15 64/67 [===========================>..] - ETA: 0s - loss: 0.1292 - accuracy: 0.9521 Epoch 9: val_loss improved from 0.01389 to 0.01280, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.1297 - accuracy: 0.9518 - val_loss: 0.0128 - val_accuracy: 1.0000 Epoch 10/15 65/67 [============================>.] - ETA: 0s - loss: 0.1198 - accuracy: 0.9572 Epoch 10: val_loss improved from 0.01280 to 0.01030, saving model to best_model.h5 67/67 [==============================] - 1s 18ms/step - loss: 0.1195 - accuracy: 0.9572 - val_loss: 0.0103 - val_accuracy: 1.0000 Epoch 11/15 65/67 [============================>.] - ETA: 0s - loss: 0.1118 - accuracy: 0.9601 Epoch 11: val_loss did not improve from 0.01030 67/67 [==============================] - 1s 17ms/step - loss: 0.1112 - accuracy: 0.9603 - val_loss: 0.1081 - val_accuracy: 0.9844 Epoch 12/15 65/67 [============================>.] - ETA: 0s - loss: 0.0998 - accuracy: 0.9652 Epoch 12: val_loss did not improve from 0.01030 67/67 [==============================] - 1s 17ms/step - loss: 0.1008 - accuracy: 0.9645 - val_loss: 0.1666 - val_accuracy: 0.9720 Epoch 13/15 65/67 [============================>.] - ETA: 0s - loss: 0.0921 - accuracy: 0.9668 Epoch 13: val_loss did not improve from 0.01030 67/67 [==============================] - 1s 17ms/step - loss: 0.0921 - accuracy: 0.9667 - val_loss: 0.0639 - val_accuracy: 0.9862 Epoch 14/15 65/67 [============================>.] - ETA: 0s - loss: 0.0907 - accuracy: 0.9677 Epoch 14: val_loss did not improve from 0.01030 67/67 [==============================] - 1s 18ms/step - loss: 0.0908 - accuracy: 0.9677 - val_loss: 0.2643 - val_accuracy: 0.9361 Epoch 15/15 65/67 [============================>.] - ETA: 0s - loss: 0.0891 - accuracy: 0.9683 Epoch 15: val_loss did not improve from 0.01030 67/67 [==============================] - 1s 18ms/step - loss: 0.0891 - accuracy: 0.9684 - val_loss: 0.1410 - val_accuracy: 0.9750 Epoch 15: early stopping
Plotting the train and validation accuracy
trainvalcurve(history_3)
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(x_test_normalized, model_3)
82/82 [==============================] - 0s 3ms/step
This model is performing realy good, and the accuracy seems to be converging.
Generate the classification report and confusion matrix
class_conf(test_labels, y_pred_test_classes)
precision recall f1-score support
0 0.97 0.96 0.97 1300
1 0.96 0.97 0.97 1300
accuracy 0.97 2600
macro avg 0.97 0.97 0.97 2600
weighted avg 0.97 0.97 0.97 2600
backend.clear_session()
from keras.preprocessing.image import ImageDataGenerator
#from tensorflow.keras.layers import RandomCrop, RandomFlip, RandomTranslation, RandomRotation, RandomZoom, RandomContrast
from sklearn.model_selection import train_test_split
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
x_train, x_val, y_train, y_val = train_test_split(train_images, y_train_encoded, test_size = val)
train_datagen = ImageDataGenerator( rotation_range=90, horizontal_flip=True )
#vertical and horizontal movement, and zoom do not produce good results. It probably has to do with the nature of the images.
val_datagen =ImageDataGenerator()
it1 = train_datagen.flow(x_train, y_train, batch_size=100, seed = 37, shuffle = True)
it2 = val_datagen.flow(x_val, y_val, batch_size=100, seed = 37, shuffle = True)
images, infected = next(it1)
cols2 = 4
rows2 = 3
fig = plt.figure(figsize = (12, 12))
for i in range(cols2):
for j in range(rows2):
ax = fig.add_subplot(rows2, cols2, i * rows2 + j + 1)
ax.imshow(images[i*rows2 + j].astype('uint8'))
if infected[i*rows2 + j,1] ==1:
ax.set_title('Infected ')
else:
ax.set_title('Uninfected ')
plt.show()
#Now lets see what happens for the same images, how this is transforming it
aux = x_train[86]
samples = np.expand_dims(aux, 0)
# prepare iterator
it = train_datagen.flow(samples, batch_size=1)
# generate samples and plot
for i in range(9):
# define subplot
plt.subplot(330 + 1 + i)
# generate batch of images
aux0 = it.next()
# convert to unsigned integers for viewing
image = aux0[0].astype('uint8')
# plot raw pixel data
plt.imshow(image)
# show the figure
plt.show()
#This is the same model as before, since it performs so well.
def model_4():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 3, padding = "same", input_shape = (size, size, 3)))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 32, kernel_size = 3, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(BatchNormalization())
model.add(Conv2D(filters = 64, kernel_size = 5, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Dropout(0.3))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_4 = model_4()
print(model_4.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 448
leaky_re_lu (LeakyReLU) (None, 80, 80, 16) 0
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
conv2d_1 (Conv2D) (None, 40, 40, 32) 4640
leaky_re_lu_1 (LeakyReLU) (None, 40, 40, 32) 0
max_pooling2d_1 (MaxPooling (None, 20, 20, 32) 0
2D)
batch_normalization (BatchN (None, 20, 20, 32) 128
ormalization)
conv2d_2 (Conv2D) (None, 20, 20, 64) 51264
leaky_re_lu_2 (LeakyReLU) (None, 20, 20, 64) 0
max_pooling2d_2 (MaxPooling (None, 10, 10, 64) 0
2D)
dropout (Dropout) (None, 10, 10, 64) 0
batch_normalization_1 (Batc (None, 10, 10, 64) 256
hNormalization)
flatten (Flatten) (None, 6400) 0
dense (Dense) (None, 32) 204832
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 261,634
Trainable params: 261,442
Non-trainable params: 192
_________________________________________________________________
None
Using Callbacks
es = EarlyStopping(monitor = 'val_loss', mode = 'min', verbose = 1, patience = 5)
mc = ModelCheckpoint('best_model.h5', monitor = 'val_accuracy', mode = 'max', verbose = 1, save_best_only = True)
Fit and Train the model
history_4 = model_4.fit(
it1,
validation_data = it2,
epochs = epoch,
batch_size=batch,
steps_per_epoch=150,
callbacks = [es, mc]
)
#This has different parameters than our previous models.For example, the validation data is a generator
# and we need to specify steps per epoch.
Epoch 1/15 150/150 [==============================] - ETA: 0s - loss: 0.6031 - accuracy: 0.6801 Epoch 1: val_accuracy improved from -inf to 0.51322, saving model to best_model.h5 150/150 [==============================] - 25s 151ms/step - loss: 0.6031 - accuracy: 0.6801 - val_loss: 1.0528 - val_accuracy: 0.5132 Epoch 2/15 150/150 [==============================] - ETA: 0s - loss: 0.3391 - accuracy: 0.8564 Epoch 2: val_accuracy improved from 0.51322 to 0.58474, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.3391 - accuracy: 0.8564 - val_loss: 1.7805 - val_accuracy: 0.5847 Epoch 3/15 150/150 [==============================] - ETA: 0s - loss: 0.1802 - accuracy: 0.9369 Epoch 3: val_accuracy improved from 0.58474 to 0.81430, saving model to best_model.h5 150/150 [==============================] - 22s 150ms/step - loss: 0.1802 - accuracy: 0.9369 - val_loss: 0.5874 - val_accuracy: 0.8143 Epoch 4/15 150/150 [==============================] - ETA: 0s - loss: 0.1358 - accuracy: 0.9530 Epoch 4: val_accuracy improved from 0.81430 to 0.96094, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.1358 - accuracy: 0.9530 - val_loss: 0.1256 - val_accuracy: 0.9609 Epoch 5/15 150/150 [==============================] - ETA: 0s - loss: 0.1144 - accuracy: 0.9614 Epoch 5: val_accuracy improved from 0.96094 to 0.96374, saving model to best_model.h5 150/150 [==============================] - 23s 151ms/step - loss: 0.1144 - accuracy: 0.9614 - val_loss: 0.0982 - val_accuracy: 0.9637 Epoch 6/15 150/150 [==============================] - ETA: 0s - loss: 0.1113 - accuracy: 0.9618 Epoch 6: val_accuracy did not improve from 0.96374 150/150 [==============================] - 22s 149ms/step - loss: 0.1113 - accuracy: 0.9618 - val_loss: 0.1788 - val_accuracy: 0.9261 Epoch 7/15 150/150 [==============================] - ETA: 0s - loss: 0.0973 - accuracy: 0.9673 Epoch 7: val_accuracy did not improve from 0.96374 150/150 [==============================] - 22s 150ms/step - loss: 0.0973 - accuracy: 0.9673 - val_loss: 0.1420 - val_accuracy: 0.9473 Epoch 8/15 150/150 [==============================] - ETA: 0s - loss: 0.0928 - accuracy: 0.9675 Epoch 8: val_accuracy improved from 0.96374 to 0.97115, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.0928 - accuracy: 0.9675 - val_loss: 0.0869 - val_accuracy: 0.9712 Epoch 9/15 150/150 [==============================] - ETA: 0s - loss: 0.0887 - accuracy: 0.9695 Epoch 9: val_accuracy improved from 0.97115 to 0.97135, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.0887 - accuracy: 0.9695 - val_loss: 0.0841 - val_accuracy: 0.9714 Epoch 10/15 150/150 [==============================] - ETA: 0s - loss: 0.0803 - accuracy: 0.9739 Epoch 10: val_accuracy improved from 0.97135 to 0.97416, saving model to best_model.h5 150/150 [==============================] - 22s 150ms/step - loss: 0.0803 - accuracy: 0.9739 - val_loss: 0.0806 - val_accuracy: 0.9742 Epoch 11/15 150/150 [==============================] - ETA: 0s - loss: 0.0746 - accuracy: 0.9748 Epoch 11: val_accuracy improved from 0.97416 to 0.97636, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.0746 - accuracy: 0.9748 - val_loss: 0.0795 - val_accuracy: 0.9764 Epoch 12/15 150/150 [==============================] - ETA: 0s - loss: 0.0783 - accuracy: 0.9747 Epoch 12: val_accuracy did not improve from 0.97636 150/150 [==============================] - 22s 150ms/step - loss: 0.0783 - accuracy: 0.9747 - val_loss: 0.1069 - val_accuracy: 0.9661 Epoch 13/15 150/150 [==============================] - ETA: 0s - loss: 0.0719 - accuracy: 0.9757 Epoch 13: val_accuracy did not improve from 0.97636 150/150 [==============================] - 23s 150ms/step - loss: 0.0719 - accuracy: 0.9757 - val_loss: 0.0737 - val_accuracy: 0.9762 Epoch 14/15 150/150 [==============================] - ETA: 0s - loss: 0.0685 - accuracy: 0.9767 Epoch 14: val_accuracy improved from 0.97636 to 0.97696, saving model to best_model.h5 150/150 [==============================] - 22s 149ms/step - loss: 0.0685 - accuracy: 0.9767 - val_loss: 0.0690 - val_accuracy: 0.9770 Epoch 15/15 150/150 [==============================] - ETA: 0s - loss: 0.0715 - accuracy: 0.9760 Epoch 15: val_accuracy did not improve from 0.97696 150/150 [==============================] - 22s 149ms/step - loss: 0.0715 - accuracy: 0.9760 - val_loss: 0.0721 - val_accuracy: 0.9766
Plot the train and validation accuracy
trainvalcurve(history_4)
Plotting the classification report and confusion matrix
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(test_images, model_4)
class_conf(test_labels, y_pred_test_classes)
82/82 [==============================] - 0s 2ms/step
precision recall f1-score support
0 0.98 0.98 0.98 1300
1 0.98 0.98 0.98 1300
accuracy 0.98 2600
macro avg 0.98 0.98 0.98 2600
weighted avg 0.98 0.98 0.98 2600
Now, let us try to use a pretrained model like VGG16 and check how it performs on our data.
backend.clear_session()
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
from tensorflow.keras import Model
from tensorflow.keras.applications.vgg16 import VGG16
vgg_model = VGG16(weights = 'imagenet',
include_top = False,
input_shape = (size, size, 3), pooling = 'max')
vgg_model.summary()
Model: "vgg16"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 80, 80, 3)] 0
block1_conv1 (Conv2D) (None, 80, 80, 64) 1792
block1_conv2 (Conv2D) (None, 80, 80, 64) 36928
block1_pool (MaxPooling2D) (None, 40, 40, 64) 0
block2_conv1 (Conv2D) (None, 40, 40, 128) 73856
block2_conv2 (Conv2D) (None, 40, 40, 128) 147584
block2_pool (MaxPooling2D) (None, 20, 20, 128) 0
block3_conv1 (Conv2D) (None, 20, 20, 256) 295168
block3_conv2 (Conv2D) (None, 20, 20, 256) 590080
block3_conv3 (Conv2D) (None, 20, 20, 256) 590080
block3_pool (MaxPooling2D) (None, 10, 10, 256) 0
block4_conv1 (Conv2D) (None, 10, 10, 512) 1180160
block4_conv2 (Conv2D) (None, 10, 10, 512) 2359808
block4_conv3 (Conv2D) (None, 10, 10, 512) 2359808
block4_pool (MaxPooling2D) (None, 5, 5, 512) 0
block5_conv1 (Conv2D) (None, 5, 5, 512) 2359808
block5_conv2 (Conv2D) (None, 5, 5, 512) 2359808
block5_conv3 (Conv2D) (None, 5, 5, 512) 2359808
block5_pool (MaxPooling2D) (None, 2, 2, 512) 0
global_max_pooling2d (Globa (None, 512) 0
lMaxPooling2D)
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________
transfer_layer = vgg_model.get_layer('block3_pool')
vgg_model.trainable = False
# I don't want to re train the VGG layers.
# Add classification layers on top of it
x = Flatten()(transfer_layer.output)
x = Dropout(0.3)(x)
x = Dense(32, activation = 'relu')(x)
x = BatchNormalization()(x)
pred = Dense(2, activation = 'softmax')(x)
# Initializing the model
model_5 = Model(vgg_model.input, pred)
model_5.compile(loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics = ['accuracy'])
using callbacks
es = EarlyStopping(monitor = 'val_loss', mode = 'min', verbose = 1, patience = 5)
mc = ModelCheckpoint('best_model.h5', monitor = 'val_accuracy', mode = 'max', verbose = 1, save_best_only = True)
model_5.summary()
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None, 80, 80, 3)] 0
block1_conv1 (Conv2D) (None, 80, 80, 64) 1792
block1_conv2 (Conv2D) (None, 80, 80, 64) 36928
block1_pool (MaxPooling2D) (None, 40, 40, 64) 0
block2_conv1 (Conv2D) (None, 40, 40, 128) 73856
block2_conv2 (Conv2D) (None, 40, 40, 128) 147584
block2_pool (MaxPooling2D) (None, 20, 20, 128) 0
block3_conv1 (Conv2D) (None, 20, 20, 256) 295168
block3_conv2 (Conv2D) (None, 20, 20, 256) 590080
block3_conv3 (Conv2D) (None, 20, 20, 256) 590080
block3_pool (MaxPooling2D) (None, 10, 10, 256) 0
flatten (Flatten) (None, 25600) 0
dropout (Dropout) (None, 25600) 0
dense (Dense) (None, 32) 819232
batch_normalization (BatchN (None, 32) 128
ormalization)
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 2,554,914
Trainable params: 819,362
Non-trainable params: 1,735,552
_________________________________________________________________
Fit and Train the model
history_5 = model_5.fit(
x_train_normalized, y_train_encoded,
epochs = epoch,
validation_split = val,
shuffle = True,
batch_size=batch,
use_multiprocessing = True,
callbacks = [es, mc]
)
Epoch 1/15 67/67 [==============================] - ETA: 0s - loss: 0.3706 - accuracy: 0.8493 Epoch 1: val_accuracy improved from -inf to 0.70092, saving model to best_model.h5 67/67 [==============================] - 6s 55ms/step - loss: 0.3706 - accuracy: 0.8493 - val_loss: 0.7094 - val_accuracy: 0.7009 Epoch 2/15 66/67 [============================>.] - ETA: 0s - loss: 0.2247 - accuracy: 0.9264 Epoch 2: val_accuracy improved from 0.70092 to 0.89804, saving model to best_model.h5 67/67 [==============================] - 2s 31ms/step - loss: 0.2246 - accuracy: 0.9265 - val_loss: 0.2882 - val_accuracy: 0.8980 Epoch 3/15 65/67 [============================>.] - ETA: 0s - loss: 0.1730 - accuracy: 0.9465 Epoch 3: val_accuracy improved from 0.89804 to 0.95232, saving model to best_model.h5 67/67 [==============================] - 2s 31ms/step - loss: 0.1726 - accuracy: 0.9464 - val_loss: 0.1250 - val_accuracy: 0.9523 Epoch 4/15 67/67 [==============================] - ETA: 0s - loss: 0.1424 - accuracy: 0.9553 Epoch 4: val_accuracy did not improve from 0.95232 67/67 [==============================] - 2s 30ms/step - loss: 0.1424 - accuracy: 0.9553 - val_loss: 0.1630 - val_accuracy: 0.9363 Epoch 5/15 66/67 [============================>.] - ETA: 0s - loss: 0.1273 - accuracy: 0.9581 Epoch 5: val_accuracy did not improve from 0.95232 67/67 [==============================] - 2s 30ms/step - loss: 0.1275 - accuracy: 0.9579 - val_loss: 0.1436 - val_accuracy: 0.9475 Epoch 6/15 65/67 [============================>.] - ETA: 0s - loss: 0.1079 - accuracy: 0.9652 Epoch 6: val_accuracy improved from 0.95232 to 0.96214, saving model to best_model.h5 67/67 [==============================] - 2s 31ms/step - loss: 0.1073 - accuracy: 0.9652 - val_loss: 0.0994 - val_accuracy: 0.9621 Epoch 7/15 65/67 [============================>.] - ETA: 0s - loss: 0.1017 - accuracy: 0.9674 Epoch 7: val_accuracy improved from 0.96214 to 0.97516, saving model to best_model.h5 67/67 [==============================] - 2s 31ms/step - loss: 0.1018 - accuracy: 0.9672 - val_loss: 0.0712 - val_accuracy: 0.9752 Epoch 8/15 65/67 [============================>.] - ETA: 0s - loss: 0.0961 - accuracy: 0.9692 Epoch 8: val_accuracy did not improve from 0.97516 67/67 [==============================] - 2s 30ms/step - loss: 0.0957 - accuracy: 0.9693 - val_loss: 0.1421 - val_accuracy: 0.9463 Epoch 9/15 65/67 [============================>.] - ETA: 0s - loss: 0.0865 - accuracy: 0.9716 Epoch 9: val_accuracy did not improve from 0.97516 67/67 [==============================] - 2s 30ms/step - loss: 0.0867 - accuracy: 0.9716 - val_loss: 0.2092 - val_accuracy: 0.9239 Epoch 10/15 66/67 [============================>.] - ETA: 0s - loss: 0.0837 - accuracy: 0.9730 Epoch 10: val_accuracy improved from 0.97516 to 0.97937, saving model to best_model.h5 67/67 [==============================] - 2s 31ms/step - loss: 0.0834 - accuracy: 0.9732 - val_loss: 0.0593 - val_accuracy: 0.9794 Epoch 11/15 65/67 [============================>.] - ETA: 0s - loss: 0.0777 - accuracy: 0.9734 Epoch 11: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 30ms/step - loss: 0.0778 - accuracy: 0.9735 - val_loss: 0.1051 - val_accuracy: 0.9583 Epoch 12/15 66/67 [============================>.] - ETA: 0s - loss: 0.0729 - accuracy: 0.9746 Epoch 12: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 30ms/step - loss: 0.0734 - accuracy: 0.9744 - val_loss: 0.1539 - val_accuracy: 0.9453 Epoch 13/15 65/67 [============================>.] - ETA: 0s - loss: 0.0722 - accuracy: 0.9742 Epoch 13: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 30ms/step - loss: 0.0723 - accuracy: 0.9741 - val_loss: 0.4048 - val_accuracy: 0.8353 Epoch 14/15 65/67 [============================>.] - ETA: 0s - loss: 0.0651 - accuracy: 0.9774 Epoch 14: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 30ms/step - loss: 0.0648 - accuracy: 0.9776 - val_loss: 0.4181 - val_accuracy: 0.8371 Epoch 15/15 67/67 [==============================] - ETA: 0s - loss: 0.0663 - accuracy: 0.9761 Epoch 15: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 30ms/step - loss: 0.0663 - accuracy: 0.9761 - val_loss: 0.1447 - val_accuracy: 0.9431 Epoch 15: early stopping
Plot the train and validation accuracy
trainvalcurve(history_5)
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(x_test_normalized, model_5)
82/82 [==============================] - 1s 5ms/step
Plotting the classification report and confusion matrix
class_conf(test_labels, y_pred_test_classes)
precision recall f1-score support
0 0.98 0.94 0.96 1300
1 0.94 0.98 0.96 1300
accuracy 0.96 2600
macro avg 0.96 0.96 0.96 2600
weighted avg 0.96 0.96 0.96 2600
backend.clear_session()
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
train_hgv = np.empty((train_images.shape[0], size,size,3), dtype=int)
for x in range(train_images.shape[0]):
train_hgv[x] = cv2.cvtColor(train_images[x], cv2.COLOR_RGB2HSV)
test_hgv = np.empty((test_images.shape[0], size,size,3), dtype=int)
for x in range(test_images.shape[0]):
test_hgv[x] = cv2.cvtColor(test_images[x], cv2.COLOR_RGB2HSV)
def model_6():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 3, padding = "same", input_shape = (size, size, 3)))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 32, kernel_size = 3, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(BatchNormalization())
model.add(Conv2D(filters = 64, kernel_size = 5, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Dropout(0.3))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_6 = model_6()
print(model_6.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 448
leaky_re_lu (LeakyReLU) (None, 80, 80, 16) 0
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
conv2d_1 (Conv2D) (None, 40, 40, 32) 4640
leaky_re_lu_1 (LeakyReLU) (None, 40, 40, 32) 0
max_pooling2d_1 (MaxPooling (None, 20, 20, 32) 0
2D)
batch_normalization (BatchN (None, 20, 20, 32) 128
ormalization)
conv2d_2 (Conv2D) (None, 20, 20, 64) 51264
leaky_re_lu_2 (LeakyReLU) (None, 20, 20, 64) 0
max_pooling2d_2 (MaxPooling (None, 10, 10, 64) 0
2D)
dropout (Dropout) (None, 10, 10, 64) 0
batch_normalization_1 (Batc (None, 10, 10, 64) 256
hNormalization)
flatten (Flatten) (None, 6400) 0
dense (Dense) (None, 32) 204832
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 261,634
Trainable params: 261,442
Non-trainable params: 192
_________________________________________________________________
None
history_6 = model_6.fit(
train_hgv, y_train_encoded,
epochs = epoch,
validation_split = val,
shuffle = True,
batch_size=batch,
use_multiprocessing = True,
callbacks = [es, mc]
)
Epoch 1/15 66/67 [============================>.] - ETA: 0s - loss: 0.5393 - accuracy: 0.7366 Epoch 1: val_accuracy did not improve from 0.97937 67/67 [==============================] - 5s 35ms/step - loss: 0.5386 - accuracy: 0.7377 - val_loss: 8.6740 - val_accuracy: 0.0000e+00 Epoch 2/15 66/67 [============================>.] - ETA: 0s - loss: 0.2369 - accuracy: 0.9143 Epoch 2: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 25ms/step - loss: 0.2363 - accuracy: 0.9146 - val_loss: 3.8663 - val_accuracy: 0.0088 Epoch 3/15 67/67 [==============================] - ETA: 0s - loss: 0.1379 - accuracy: 0.9530 Epoch 3: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 24ms/step - loss: 0.1379 - accuracy: 0.9530 - val_loss: 0.3300 - val_accuracy: 0.8938 Epoch 4/15 67/67 [==============================] - ETA: 0s - loss: 0.1113 - accuracy: 0.9613 Epoch 4: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 24ms/step - loss: 0.1113 - accuracy: 0.9613 - val_loss: 0.1608 - val_accuracy: 0.9655 Epoch 5/15 67/67 [==============================] - ETA: 0s - loss: 0.1002 - accuracy: 0.9660 Epoch 5: val_accuracy did not improve from 0.97937 67/67 [==============================] - 2s 24ms/step - loss: 0.1002 - accuracy: 0.9660 - val_loss: 0.0844 - val_accuracy: 0.9764 Epoch 6/15 67/67 [==============================] - ETA: 0s - loss: 0.0891 - accuracy: 0.9694 Epoch 6: val_accuracy improved from 0.97937 to 0.99499, saving model to best_model.h5 67/67 [==============================] - 2s 26ms/step - loss: 0.0891 - accuracy: 0.9694 - val_loss: 0.0339 - val_accuracy: 0.9950 Epoch 7/15 67/67 [==============================] - ETA: 0s - loss: 0.0813 - accuracy: 0.9713 Epoch 7: val_accuracy did not improve from 0.99499 67/67 [==============================] - 2s 24ms/step - loss: 0.0813 - accuracy: 0.9713 - val_loss: 0.0785 - val_accuracy: 0.9826 Epoch 8/15 67/67 [==============================] - ETA: 0s - loss: 0.0761 - accuracy: 0.9743 Epoch 8: val_accuracy did not improve from 0.99499 67/67 [==============================] - 2s 25ms/step - loss: 0.0761 - accuracy: 0.9743 - val_loss: 0.0399 - val_accuracy: 0.9902 Epoch 9/15 66/67 [============================>.] - ETA: 0s - loss: 0.0733 - accuracy: 0.9745 Epoch 9: val_accuracy did not improve from 0.99499 67/67 [==============================] - 2s 25ms/step - loss: 0.0730 - accuracy: 0.9747 - val_loss: 0.0884 - val_accuracy: 0.9760 Epoch 10/15 67/67 [==============================] - ETA: 0s - loss: 0.0662 - accuracy: 0.9769 Epoch 10: val_accuracy did not improve from 0.99499 67/67 [==============================] - 2s 25ms/step - loss: 0.0662 - accuracy: 0.9769 - val_loss: 0.0455 - val_accuracy: 0.9874 Epoch 11/15 66/67 [============================>.] - ETA: 0s - loss: 0.0630 - accuracy: 0.9782 Epoch 11: val_accuracy did not improve from 0.99499 67/67 [==============================] - 2s 25ms/step - loss: 0.0627 - accuracy: 0.9783 - val_loss: 0.0513 - val_accuracy: 0.9892 Epoch 11: early stopping
trainvalcurve(history_6)
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(test_hgv, model_6)
82/82 [==============================] - 0s 3ms/step
class_conf(test_labels, y_pred_test_classes)
precision recall f1-score support
0 0.98 0.98 0.98 1300
1 0.98 0.98 0.98 1300
accuracy 0.98 2600
macro avg 0.98 0.98 0.98 2600
weighted avg 0.98 0.98 0.98 2600
Final Model
For the final model I use HSV images and Data Augmentation, with the same architecture used before.
backend.clear_session()
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
x_train, x_val, y_train, y_val = train_test_split(train_hgv, y_train_encoded, test_size = val)
train_datagen = ImageDataGenerator( rotation_range=90, horizontal_flip=True )
#vertical and horizontal movement, and zoom do not produce good results. It probably has to do with the nature of the images.
val_datagen =ImageDataGenerator()
it1 = train_datagen.flow(x_train, y_train, batch_size=100, seed = 37, shuffle = True)
it2 = val_datagen.flow(x_val, y_val, batch_size=100, seed = 37, shuffle = True)
def model_7():
model = Sequential()
model.add(Conv2D(filters = 16, kernel_size = 3, padding = "same", input_shape = (size, size, 3)))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Conv2D(filters = 32, kernel_size = 3, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 4))
model.add(BatchNormalization())
model.add(Conv2D(filters = 64, kernel_size = 5, padding = 'same'))
model.add(LeakyReLU(0.1))
model.add(MaxPooling2D(pool_size = 2))
model.add(Dropout(0.3))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(32, activation = 'relu'))
model.add(Dense(2, activation = 'softmax'))
model.compile(
loss = 'binary_crossentropy',
optimizer = tf.keras.optimizers.Adamax(learning_rate = 0.001),
metrics=['accuracy']
)
return model
model_7 = model_7()
print(model_7.summary())
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 80, 80, 16) 448
leaky_re_lu (LeakyReLU) (None, 80, 80, 16) 0
max_pooling2d (MaxPooling2D (None, 40, 40, 16) 0
)
conv2d_1 (Conv2D) (None, 40, 40, 32) 4640
leaky_re_lu_1 (LeakyReLU) (None, 40, 40, 32) 0
max_pooling2d_1 (MaxPooling (None, 10, 10, 32) 0
2D)
batch_normalization (BatchN (None, 10, 10, 32) 128
ormalization)
conv2d_2 (Conv2D) (None, 10, 10, 64) 51264
leaky_re_lu_2 (LeakyReLU) (None, 10, 10, 64) 0
max_pooling2d_2 (MaxPooling (None, 5, 5, 64) 0
2D)
dropout (Dropout) (None, 5, 5, 64) 0
batch_normalization_1 (Batc (None, 5, 5, 64) 256
hNormalization)
flatten (Flatten) (None, 1600) 0
dense (Dense) (None, 32) 51232
dense_1 (Dense) (None, 2) 66
=================================================================
Total params: 108,034
Trainable params: 107,842
Non-trainable params: 192
_________________________________________________________________
None
np.random.seed(37)
random.seed(37)
tf.random.set_seed(37)
history_7 = model_7.fit(
it1,
validation_data = it2,
epochs = epoch,
batch_size=batch,
steps_per_epoch=150,
callbacks = [es, mc]
)
Epoch 1/15 150/150 [==============================] - ETA: 0s - loss: 0.3227 - accuracy: 0.8623 Epoch 1: val_accuracy improved from -inf to 0.60196, saving model to best_model.h5 150/150 [==============================] - 35s 161ms/step - loss: 0.3227 - accuracy: 0.8623 - val_loss: 0.7209 - val_accuracy: 0.6020 Epoch 2/15 150/150 [==============================] - ETA: 0s - loss: 0.1334 - accuracy: 0.9558 Epoch 2: val_accuracy improved from 0.60196 to 0.96434, saving model to best_model.h5 150/150 [==============================] - 24s 162ms/step - loss: 0.1334 - accuracy: 0.9558 - val_loss: 0.1077 - val_accuracy: 0.9643 Epoch 3/15 150/150 [==============================] - ETA: 0s - loss: 0.1107 - accuracy: 0.9633 Epoch 3: val_accuracy did not improve from 0.96434 150/150 [==============================] - 24s 161ms/step - loss: 0.1107 - accuracy: 0.9633 - val_loss: 0.1019 - val_accuracy: 0.9641 Epoch 4/15 150/150 [==============================] - ETA: 0s - loss: 0.0999 - accuracy: 0.9673 Epoch 4: val_accuracy did not improve from 0.96434 150/150 [==============================] - 24s 160ms/step - loss: 0.0999 - accuracy: 0.9673 - val_loss: 0.1612 - val_accuracy: 0.9519 Epoch 5/15 150/150 [==============================] - ETA: 0s - loss: 0.0932 - accuracy: 0.9697 Epoch 5: val_accuracy improved from 0.96434 to 0.96474, saving model to best_model.h5 150/150 [==============================] - 24s 162ms/step - loss: 0.0932 - accuracy: 0.9697 - val_loss: 0.1033 - val_accuracy: 0.9647 Epoch 6/15 150/150 [==============================] - ETA: 0s - loss: 0.0872 - accuracy: 0.9715 Epoch 6: val_accuracy did not improve from 0.96474 150/150 [==============================] - 24s 160ms/step - loss: 0.0872 - accuracy: 0.9715 - val_loss: 0.1122 - val_accuracy: 0.9607 Epoch 7/15 150/150 [==============================] - ETA: 0s - loss: 0.0821 - accuracy: 0.9730 Epoch 7: val_accuracy improved from 0.96474 to 0.96975, saving model to best_model.h5 150/150 [==============================] - 24s 161ms/step - loss: 0.0821 - accuracy: 0.9730 - val_loss: 0.0842 - val_accuracy: 0.9698 Epoch 8/15 150/150 [==============================] - ETA: 0s - loss: 0.0774 - accuracy: 0.9743 Epoch 8: val_accuracy improved from 0.96975 to 0.97055, saving model to best_model.h5 150/150 [==============================] - 24s 159ms/step - loss: 0.0774 - accuracy: 0.9743 - val_loss: 0.0844 - val_accuracy: 0.9706 Epoch 9/15 150/150 [==============================] - ETA: 0s - loss: 0.0730 - accuracy: 0.9755 Epoch 9: val_accuracy improved from 0.97055 to 0.97296, saving model to best_model.h5 150/150 [==============================] - 24s 160ms/step - loss: 0.0730 - accuracy: 0.9755 - val_loss: 0.0752 - val_accuracy: 0.9730 Epoch 10/15 150/150 [==============================] - ETA: 0s - loss: 0.0743 - accuracy: 0.9753 Epoch 10: val_accuracy improved from 0.97296 to 0.97556, saving model to best_model.h5 150/150 [==============================] - 24s 161ms/step - loss: 0.0743 - accuracy: 0.9753 - val_loss: 0.0692 - val_accuracy: 0.9756 Epoch 11/15 150/150 [==============================] - ETA: 0s - loss: 0.0706 - accuracy: 0.9764 Epoch 11: val_accuracy did not improve from 0.97556 150/150 [==============================] - 25s 164ms/step - loss: 0.0706 - accuracy: 0.9764 - val_loss: 0.0785 - val_accuracy: 0.9712 Epoch 12/15 150/150 [==============================] - ETA: 0s - loss: 0.0690 - accuracy: 0.9773 Epoch 12: val_accuracy improved from 0.97556 to 0.97897, saving model to best_model.h5 150/150 [==============================] - 24s 163ms/step - loss: 0.0690 - accuracy: 0.9773 - val_loss: 0.0675 - val_accuracy: 0.9790 Epoch 13/15 150/150 [==============================] - ETA: 0s - loss: 0.0674 - accuracy: 0.9778 Epoch 13: val_accuracy did not improve from 0.97897 150/150 [==============================] - 24s 161ms/step - loss: 0.0674 - accuracy: 0.9778 - val_loss: 0.0661 - val_accuracy: 0.9752 Epoch 14/15 150/150 [==============================] - ETA: 0s - loss: 0.0669 - accuracy: 0.9781 Epoch 14: val_accuracy did not improve from 0.97897 150/150 [==============================] - 24s 160ms/step - loss: 0.0669 - accuracy: 0.9781 - val_loss: 0.0771 - val_accuracy: 0.9722 Epoch 15/15 150/150 [==============================] - ETA: 0s - loss: 0.0595 - accuracy: 0.9807 Epoch 15: val_accuracy did not improve from 0.97897 150/150 [==============================] - 24s 159ms/step - loss: 0.0595 - accuracy: 0.9807 - val_loss: 0.0678 - val_accuracy: 0.9786
trainvalcurve(history_7)
y_pred_test, y_pred_test_classes, y_pred_test_max_probas = evaltest(test_hgv, model_7)
class_conf(test_labels, y_pred_test_classes)
82/82 [==============================] - 0s 4ms/step
precision recall f1-score support
0 0.99 0.98 0.98 1300
1 0.98 0.99 0.98 1300
accuracy 0.98 2600
macro avg 0.98 0.98 0.98 2600
weighted avg 0.98 0.98 0.98 2600
Comments:
The recall improves slightly from the previous model. Now we are able to identify correctly 99% of the infected cells.
This is probably due to having more similarities between training and test samples when we do data augmentation, since it allows for more variations of the images.
Now, I show some of the infected cells the model is not detecting, both as HSV and RGB images.
cols2 = 4
rows2 = 3
fig = plt.figure(figsize = (12, 12))
for i in range(cols2):
for j in range(rows2):
ax = fig.add_subplot(rows2, cols2, i * rows2 + j + 1)
ax.imshow(aux[i*rows2 + j].astype('uint8'))
plt.show()
aux = test_images[np.logical_and(test_labels == 1, y_pred_test_classes ==0) ]
cols2 = 4
rows2 = 3
fig = plt.figure(figsize = (12, 12))
for i in range(cols2):
for j in range(rows2):
ax = fig.add_subplot(rows2, cols2, i * rows2 + j + 1)
ax.imshow(aux[i*rows2 + j].astype('uint8'))
plt.show()
The images show 12 out of 14 infected cells not detected. It looks like the model is not good at detecting infections that are close to the black part, and some that are more subtle. Given more time this problem could be addressed by modifying the architecture accordingly.
The most meanigful insights from the data are that CNN seems to perform very good.
Also, from playing around, it looks that some data augmentation techiques are not useful. Thes probably has to do with the fact that photos are taken in a controlled environment. Therefore, including zooms or horizontal/vertical movements agreggate noise to the images and make the techniques perform worse. This is important since in a real life situation we would need to prepare the images in a way it ressembles the ones used here.
Afrter playing around with different resolutions, more complicated techniques, epochs, etc., it seems that there are little gains from doing that.
As mentioned before, it seems there is little gain from complicating the model further. The performance is quite good with only 3 convolution layers.
More over, even though there are many values that have to be calibrated, as batch, size of the image, epoch, etc. Results (not shown here) do not vary much.
In this case, using a simple model with HSV images and Data Augmentation gives the best performance, but is not far from the results with only HSV, or with standarized RGB and Data Augmentation. Again, this is probably due to the nature of the problem and how the input data was prepared.